pablo rodriguez telefonica research - · pdf filecentralize what you can, distribute what...

Centralize what you can, Distribute what you must

Pablo Rodriguez Telefonica Research

© 2010 Telefónica Investigación y Desarrollo, S.A. Unipersonal

Distributed Architectures

  Last decade networking research has focused in highly distributed architectures

  P2P, Distributed Networking Systems

  High availability, reliability, high performance


End Result

  The end result is the commoditization of Hardware equipment and reduction of related costs (CAPEX)

—  Distributed Cloud servers, Cheaper databases, Commoditizing routers, etc

  However, operational costs (OPEX), have not decreased at the same rate

  In fact, highly distributed architectures tend to increase operational costs

—  Human costs are harder to lower than technology costs


Operational Costs

  Usually account for a much higher cost than the actual equipment

  As CAPEX comes down, OPEX still remains very high

  the CLOUD helps replace CAPEX for OPEX and lower OPEX costs due to higher economies of scale

  Centralize what you can and distributed what you must —  absorb traffic early, reliability, responsiveness, latency, etc.


Cloud Benefits

  Reduces OPEX

  Do it once and share

  Do it more coordinately

  Do it more efficiently

5


sample benefits for mobile cloud applications: more than offloading computation

6


Impact of Smart Phones on the network

7

  Signaling Problems: Fast Dormancy for battery savings

1

CELL_DCH: Dedicated Channel CELL_FACH: Connected mode but Shared Channel CELL_PCH: Mostly inactive radio. Radio Resource Control connection up. Idle: Mostly inactive radio. no Radio Resource Control connection.

DCH

FACH

PCH

DCH

Idle



8

  Signaling Problems

1



  Massive rise in the amount of signalling on Radio and Core systems due to both chatty applications and device designs

  Saturated network elements (e.g. firewalls), due to uncoordinated connection management

  High number of unnecessary PPP context setup and releases

9

1


Can we run a full/partial virtual instance of each smartphone in the cloud (apps/firmware?)

- proxy a lot of the functionality, to save signaling and mitigate network impact - share info and measurements across devices, - coordinate device sync ups to avoid network peaks, wasting battery, or interfering with voice traffic

10

1

Application servers

Mobile cloud intelligence


Tput Transfer (J/MB)

Idle (W) Scan (W) Range (m)

Cellular Few kbps 100

100 0* 0* 500

Wi-Fi 11-54 Mbps 5 0.77 1.29 100 Bluetooth 700 Kbps 0.1 0.01 0.12 10

Battery efficient Ad-Hoc Data Sharing

  By sharing location info and data content across devices using the cloud, one can predict nearby device and data presence better, coordinate short range local transfers and save energy

11

Ananthanarayan et al.’09

2


Battery efficient Ad-Hoc Data Sharing

  Centrally coordinating WiFi/Bluetooth wake up intervals, one can save energy and better predict sharing opportunities: Global View. Local view may see a disconnected graph, while global view may be able to see a connected one

12

1) Location

1) Location

3) Short range data exchange

2) Instruction to send data to another user

2


Summary

  As hardware is more commoditized, Operational costs will dominate

  Cloud reduces OPEX by amortizing support/management costs over large number of applications

  Cloud also provides more intelligent services by aggregating global information and intelligently coordinating distributed actions

  What to centralize, what to leave distributed? That is the question

13

pablo rodriguez telefonica research - · pdf filecentralize what you can, distribute what...

Documents